AI’s New Moravec Paradox Sparks Debate
Recent posts converge on the same frustration: today’s LLMs can appear brilliant on highly abstract problems, yet remain unreliable on ordinary real-world work. The current takeaway from this discussion is not that models are weak at reasoning in general, but that their strengths are uneven and often poorly aligned with practical business tasks. That matters because recent progress in math or coding does not automatically mean agents are ready for messy, multi-step workflows.
Confirmed
- @YuchenjUW summarized the contrast bluntly: AI may solve around 10 open math problems, yet still fail to follow 3 simple instructions, while wasting tens of millions of tokens and hundreds of dollars in the process.
- @shakoistsLog pointed to a concrete business example: AI agents can perform strongly on advanced math, but still make mistakes on something as mundane as editing a customer invoice. In their framing, the gap is between abstract reasoning and the dense, exception-filled rules of the real world.
- @rickasaurus made the same point through taxes: models may be heading toward solving problems as hard as the Riemann Hypothesis, yet still cannot be trusted to handle everyday tax work.
- @NYCounihan reported a similar pattern in programming. Even if a model looks impressive on difficult problems, it may drift off course during a 5-hour coding session and generate useless side quests instead of staying focused on the main task.
- In a reposted remark, security researcher halvarflake joked that spending thousands of dollars of compute to solve a Fields Medal-level math conjecture would be less impressive if the same systems still cannot fix his everyday bugs.
Why it matters
- @paraschopra described this as a “new Moravec paradox” for the LLM era: proving a difficult mathematical conjecture may be easier for a model than building a simple app that earns $100 on the App Store. His point is that benchmark-style brilliance does not cleanly translate into economic usefulness.
- @ajratner argued that this is a modern replay of Moravec’s paradox: tasks humans see as difficult, such as math and programming, can become relatively tractable for AI, while supposedly simple daily tasks remain hard. He warned against linearly extrapolating from progress in domains with strong verification signals to domains that lack rich feedback and clear ground truth.
- @Franc0Fernand0 suggested the key gap is not raw intelligence but the ability to notice contradictions and ask clarifying questions. In his comparison, a human intern will stop and ask when something looks off, while current AI often proceeds anyway, which helps explain failures in real operational settings.
2026-08-02 ~ 2026-08-03 · 8 related posts
Primary sources
- [source] AI Agents Ace Complex Math But Struggle to Update Invoices — shakoistsLog · 2026-08-02
- AI Will Solve the Riemann Hypothesis Before Doing Our Taxes — rickasaurus · 2026-08-02
- The Irony of AI: Solves Open Math Problems, Fails 3 Simple Instructions — Yuchenj_UW · 2026-08-03
- AI Solves Fields Medal Math But Fails Routine Debugging, Expert Jokes — mboehme_ · 2026-08-03
- AI's New Moravec's Paradox: Math and Coding Soar While Other Domains Lag — ajratner · 2026-08-03
- [source] AI Solves Math Problems but Fails Simple Instructions? The Gap is About Asking Questions — Franc0Fernand0 · 2026-08-03
- AI Can Solve Math Equations But Wanders Off During 5-Hour Coding Tasks — NYCounihan · 2026-08-03
- [source] New Moravec's paradox: LLMs solve math but struggle to build a $100 app — paraschopra · 2026-08-03